Paper puts an LLM in charge of choosing the search space for autonomous labs
bravo_abad · x · 2026-08-26
Self-driving labs today are autonomous only inside a search space that a human defines. A new paper by Yuma Iwasaki and coauthors automates that decision too:
- A hierarchical materials-discovery framework where an LLM operates outside the conventional Bayesian-optimization loop, repeatedly deciding which materials space to explore next;
- Within each space, Gaussian-process Bayesian optimization picks the next composition to evaluate;
- Once exhausted, an outer loop analyzes accumulated results — the LLM receives measured compositions and target values as a CSV and proposes the next three-element search space.
The distinction matters: optimizing how to search was already automated; where to search is now delegated to the LLM as well.
More from Research
- EMNLP paper DemoICL: learning novel procedural knowledge from in-context video demos — liuziwei7 · 2026-08-26
- Compressing capable representations is easier than enhancing compressed ones — antoine_chaffin · 2026-08-26
- V-GIFT Boosts Visual Reasoning via Self-Supervised Guidance — srchvrs · 2026-08-26
- GigaBrain-0.7: Three-System Architecture Improves Embodied Generalization — open-gigaai · 2026-08-26
- LSTM by Hand: A 15-step walkthrough of calculation — ProfTomYeh · 2026-08-26
- AI for chip design is underrated; OpenAI explores hardware-in-the-loop — ai · 2026-08-26